Predicting Pipe Breaks

in the City of Atlanta

​

Thach Tran

Joseph Moravitz

Shweta Shalini

Anum Alimohammed

Portia Essuman

Presentation Outline

1. The Problem--An Aging Water Distribution Infrastructure
2. Why it Matters--The Cost of Main Breaks
3. Our Solution -- Applying Machine Learning to a Real-World Problem

The Problem

City of Atlanta’s Water Distribution

• System dates back to 1875
• 2,790 miles of water mains
• 246 million gallons per day
• 1.2 million people per day

​

​

America's Aging Water Infrastructure

• Main breaks are becoming more common due to an aging infrastructure.
• About 240,000 main breaks/year!
• It will cost up to $1 trillion dollars over the next 25 years to maintain service and meet demands
• 1.7 trillion gallons of water lost per year at a national cost of $2.6 billion per year

​

Pipe Breaks in ATLANTA are an Issue

• The city’s water pipe infrastructure is aging and in disrepair
• A 2017 audit found that the city loses as much as 30% of the water produced due to ‘aging infrastructure’
• Ex: 1-day water outage + boil advisory in Dec 2018 linked to leaky pipes/low pressure

​

About the Data

• Dataset on water main system from the City of Atlanta’s Department of Watershed Management
• Combination of geospatial data and hydraulic modeling data
• Includes information on pipes characteristics as well as some conditions the pipes are under
• 56,000+ rows

​

​

​

​

Overview of Pipe Failures in Atlanta

Graph of Pipe Failures Over Time: 1997-2017

Trends in Pipe Failures

SEASONALITY

Trends in Pipe Failures

AGE

Trends in Pipe Failures

MATERIALS

Why does it matter?

Cost to Fix a Break

• The Atlanta Watershed Commissioner says each break costs between $6,000 and $20,000 to fix

​

Cost to Inspect a Pipe before Break

​

• On the average, pipe inspections in Atlanta cost $803. With a low and high end of $350 and $1500
• But inspecting EVERY pipe is prohibitively costly and time consuming

Existing Solutions

Existing solutions

Our Solution

GOAL:

Identify water mains that are at high risk of failure in order to prevent breaks and reduce repair costs

Our Solution

A machine learning algorithm that predicts which pipes are most likely to break

Benefits:

• No disruption to service
• Reduce Costs
• Improve Performance
• Improve Satisfaction (customer)
• Improve Safety
• Avoid Catastrophic Events/Expenses

Data Cleaning Process

• We used pandas to clean the data
• Lots of one-Hot encoding of qualitative data (columns with strings)

Machine Learning Model

• We tried many different models:
• Logistic Regression Analysis using SciKit Learn
• Neural Network using Keras
• Random Forest with Grid Searching **WE USED THIS ONE**

​

• Troubleshooting:
• Unbroken pipes overrepresented in sample
• To address this, we undersampled unbroken pipes and randomize the selection of unbroken pipes.

Model Performance

Confusion Matrix

​

​

​

​

​

Accuracy : 80%

Model Performance

Classification Report

​

Key Findings:

Factors that Affect Likelihood of a Pipe Break

We considered 57 unique factors

Top Factors--by feature importance:

• Pipe Length (0.0899)
• Age (0.0845)
• Roughness (0.038)
• Max Head (0.0375)
• Max Head Loss (0.0365)

​

​

​

​

​

Shortcomings of Our Model

• We have not included all the factors that could impact our model. For example we have data about surrounding soil type but it is too specific.
• We are forced to undersample our data to avoid bias in training the model, however this limits certain aspects of its performance.
• We wanted to explore the data more but due to limited time we could not do it.

Thank You!

Shweta wants to know. . .

“Do you wanna play in the rain of the broken water main?”